A voice-first Windows assistant with local tools and verified actions.
Talk or type, manage everyday tasks, and keep control of what runs. Built with Python, WebView2 and Gemini Live. In development — general desktop automation is still being built.
Run it · Commands · Roadmap · Changelog · Report a bug
| Area | Available now |
|---|---|
| Conversation | Live voice/text, streaming transcripts, audio replies, interruptions, Stop and Mute |
| Wake & sleep | Local wake detection, one-time sleep, return-to-sleep mode and idle sleep timer |
| Memory | Temporary conversation context and opt-in personal memory, with review or automatic saving |
| Notes & clocks | Note editing, content/date selection, batch deletion/completion, multiple timers and stopwatches |
| Dashboard | Windows system/network readings, media controls, player volume and current weather |
| Visual context | Opt-in camera or screen sharing during Live messages; bounded JPEGs, up to 1 frame/s |
| Task foundation | Fixed plans over enabled local actions, verified progress, queued work and revision-bound Stop |
“Add a note: Review the project tomorrow.”
“Mark my last two notes as done.”
“Start a five-minute timer.”
“Create a background task to add two notes: Plan A and Plan B.”
Examples illustrate intent, not required phrases. More commands and their limits →
| Now | Next | Later |
|---|---|---|
| Live assistant + local tools | Safe Continue/recovery, files, Windows/browser/visual control and learned procedures | Executable routines, external accounts/channels and provider separation |
Not available yet: general app clicking/typing, executable learned skills/routines, Google Drive/Calendar/mail, or Telegram/Discord integrations. Skill/routine editors currently store definitions only.
Follow Stage 7 → · See the architecture →
How actions are checked
flowchart LR
U[User request] --> M[Structured model tools]
M --> G[Validation and policy]
G --> T[Single task worker]
T --> R[Observed result]
The model proposes actions; the application owns execution and verification. Stop acceptance is separate from actual completion. Completed effects are preserved; uncertain outcomes are not automatically repeated. Historical receipts do not restore private arguments or grant fresh permissions.
Windows · Python 3.11+ · Microsoft Edge WebView2 Runtime
Development is tested with Python 3.14. From the repository root:
cd v1
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements-dev.lock.txt
.\.venv\Scripts\python.exe -m app.mainAlready installed? Run the last command from v1.
Enter your own Gemini API key in Settings; the Live catalog loads automatically.
Provider access, quota and possible API charges depend on your account.
Wake-model setup & troubleshooting · UI browser preview
Portable Windows demo
Build instructions for the unsigned Windows x64 demo. Extract the whole ZIP; Python is bundled. WebView2 and .NET Framework 4.8+ are required. Keys, personal data and wake-model weights are not bundled.
The existing demo predates recent source changes. It is not a signed release or certification for every PC; no binary was rebuilt for this update.
Data, permissions & practical limits
- Conversation and explicitly shared visual context go to Google. Sleeping wake detection runs locally; this is not a fully offline assistant.
- API keys use Windows Credential Manager. Notes/settings are local; the unencrypted task journal keeps fingerprints/status metadata rather than note/file/message bodies. Original tool results stay in bounded RAM. Chat context and clocks reset on exit.
- Camera/screen capture is opt-in, uses one source at a time, and stops on sleep or Disconnect. Frames are not saved locally; already-sent content cannot be recalled.
- Personal memory is sent only when enabled. Location/device permissions are managed in Settings; permission probes do not start conversation capture.
- Readings depend on Windows/player support. CPU/GPU temperatures and VRAM capacity are unavailable. Network rates are not a speed test; token totals are not billing/quota.
- No general latency, RAM or wake-accuracy guarantee. Native startup/teardown timing issues remain under investigation. Weather's public endpoint needs a terms review before commercial deployment.
Latest checks — evidence, not a production-readiness badge
The latest Stop/Disconnect slice passed 187 targeted Python tests, 11 JS suites, 35 browser groups, 11 native phases, documentation checks and type checking. Checks use synthetic transports and temporary profiles, without paid APIs or real user data.
Manual acceptance of this slice is deferred. Automated success does not replace real-account/device testing. Earlier failed attempts and fixes remain recorded.
Found a reproducible problem? Open an issue with steps and a redacted error. For code changes, start with AGENTS.md and the current plan.
Original contributions are MIT licensed. Dependencies, UI assets and model weights retain their own terms; third-party review remains open. Weather: Open-Meteo, CC BY 4.0; place data: GeoNames.